Pin CUDA 12.8 PyTorch wheels to a compatible version

[Ha1baraA11]

Summary

• Address the reported issue: torch resolves to 2.14.0 (PyPI) or 2.11.0+cu128 (download.pytorch.org/whl/cu128) depending on install path (.github/workflows/release.yml:296).

Verification

• Repository validation completed successfully.
• This validation record does not claim manual/browser verification.
• The real-browser E2E suite was not run.

Commands run:

ruby -e 'require "yaml"; YAML.parse_file(".github/workflows/release.yml"); puts "workflow YAML parses"'
python3.12 -c 'from pathlib import Path; p=Path(".github/workflows/release.yml").read_text(); s=Path("scripts/package_cuda.py").read_text(); assert "torch==2.10.0+cu128" in p; assert "torchaudio==2.10.0+cu128" in p; assert ">=2.7.0,<2.11.0" in s; assert (2,7,0) <= (2,10,0) < (2,11,0); print("CUDA pins 2.10.0+cu128 are present and satisfy >=2.7.0,<2.11.0")'

Fixes #1099